Agents, automation and teams
Automation multiplies whatever you feed it. A broken manual process, automated, is the same broken process running at three in the morning while you sleep. This chapter is about doing that on purpose.
13 ways to get it wrong
“We rolled it out to everyone on Monday. Adoption is at four percent.”
- 6.1
Automate a process nobody has written down.
WhyYou cannot automate what you cannot describe. The undocumented steps are where the judgement lives.
ResultAn agent that handles the easy sixty percent and silently mangles the rest.
- 6.2
Give the agent write access on day one.
WhyRead-only first, then suggest, then act with approval, then act alone. Skipping stages skips the learning.
ResultFour hundred emails sent before anyone noticed the loop.
- 6.3
Build with no human checkpoint.
WhyThe value of a checkpoint is not catching every error. It is catching the category of error you did not predict.
ResultErrors compound quietly until a customer reports them for you.
- 6.4
Log nothing.
WhyWhen output goes wrong you need to see the input, the prompt and the tool call. Without logs you are guessing.
ResultA bug you cannot reproduce, so you rebuild the whole thing instead.
- 6.5
Measure nothing before you start.
WhyNo baseline means no way to prove the project worked, which means no budget next quarter.
ResultA working system that gets cancelled because nobody can show the gain.
- 6.6
Chain seven agents when one prompt would do.
WhyEvery hop adds latency, cost and a new failure point. Complexity is not sophistication.
ResultA fragile system that breaks weekly and takes an engineer to restart.
- 6.7
Pick the tool first, then look for a problem.
WhyTool-led projects optimise for the demo, not the work.
ResultTwelve subscriptions and no changed outcome.
- 6.8
Train the team once, in a webinar, and never again.
WhySkill comes from doing the real work with support nearby. A recording is not support.
ResultFour percent adoption and a quiet conclusion that the staff are resistant.
- 6.9
Punish the first person who reports an AI mistake.
WhyWhatever happens to the first reporter sets the reporting rate for everyone else.
ResultSilence. The next mistake surfaces via a customer.
- 6.10
Give a shared assistant no rules about data.
WhyStaff will paste whatever is on their screen, including contracts, salaries and health data.
ResultThe regulatory incident you did not budget for.
- 6.11
Keep the prompt logic inside one person's head.
WhyUndocumented workflows leave with the person who built them.
ResultA capability that disappears with one resignation.
- 6.12
Cut the review step because the output looked fine last week.
WhyModel versions change under you. Quality that was stable on Tuesday may not be stable in the next release.
ResultSilent quality drift that you find out about from your customers.
- 6.13
Promise the board a headcount saving in month one.
WhyReal gains come from redesigning the work, and that takes a quarter or two.
ResultA cancelled programme just before it was about to pay off.
Grading your failure
Three levels. Anyone can reach bronze on a Monday morning.
The workflow runs and nobody uses the output.
The one person who built it leaves.
A customer emails you a transcript of your own agent.